Home · Datasets · WinoGrande
DATASET

WinoGrande

A large-scale commonsense reasoning dataset inspired by the Winograd Schema Challenge.

TARGET QUERY winogrande benchmark · ~2K/mo
SIZE
44K fill-in-the-blank problems
CREATOR
AI2
MODALITY
text
LICENSE
Apache 2.0
RELEASED
2019-07
OVERVIEW Updated 2026-05-17

Overview

WinoGrande is a large-scale commonsense reasoning dataset from AI2 testing pronoun resolution requiring world knowledge. Built at 44,000 problems using crowdsourcing with adversarial filtering, it provides more robust evaluation than the original Winograd Schema Challenge.

What’s In It

WinoGrande contains 44,000 fill-in-the-blank problems where the correct answer requires commonsense reasoning about physical properties, social interactions, or everyday knowledge. Adversarial filtering ensures problems cannot be solved by statistical shortcuts.

How It’s Used

WinoGrande is part of standard language model evaluation suites including the Open LLM Leaderboard. It tests whether models possess implicit world knowledge necessary for natural language understanding.

Controversies

Despite adversarial filtering, some problems may contain spurious statistical cues. The binary-choice format limits assessment to recognition rather than generation of commonsense knowledge.